When it goes wrong

Beemy will get some of these wrong.

Any system reading thousands of messages and making judgement calls will misjudge some of them. The question worth answering honestly is what one wrong call costs you, and how fast you get it back. That matters more than how often it happens.

9 min read

How the cost is capped

The price of a mistake is set by the rung, regardless of the mistake itself.

Every category of task sits on one of three rungs of the autonomy ladder. You're the one who decides which rung it's on. That decision also sets a ceiling on what a wrong call there can do.

  1. 01

    Read-only

    Worst case A message gets filed in the wrong place: sorted as routine when it mattered, or flagged when it didn't.

    Nothing leaves your account at this rung. The message is still there, still readable, and re-filing it is a one-tap correction, not damage control.

  2. 02

    Approve before it sends

    Worst case A draft that isn't quite right sits in your drafts folder, waiting for you.

    It costs you exactly the seconds it took to read and reject it. Nothing sends until you approve it, so a bad draft is a non-event, not an incident.

  3. 03

    Scoped autonomy

    Worst case An action you switched on yourself, for one category, reaches someone else before you see it.

    This is the only rung where a mistake leaves your account. It's bounded to exactly the category you turned on. Every action here is logged and reviewable after the fact. Turning the dial back down is as fast as turning it up.

The ladder itself, and how trust moves up it, is laid out in full on Trust & security. See On Your Behalf for exactly what each rung lets Beemy send without you.

The catalogue

Eight ways it goes wrong, named in the open.

Each one below is a shape a mistake actually takes. Two of them are less bugs than the questions people already hold before connecting an account. We answer them here, instead of leaving you to wonder.

Something important got triaged as routine

Looks like A message that actually mattered lands in the ordinary pile instead of getting flagged for you.

Why it happens Triage ranks a message by who it's from and how urgent it reads. That's based on the relationship history it's seen so far. A first message from someone new doesn't carry the weight an established contact would. Neither does an urgent ask with no urgent-sounding words.

What contains it This happens at the read-only rung: nothing gets sent, deleted, or acted on. The message is sitting in your inbox exactly where triage filed it, findable the moment you look.

What you do Re-file it. That's the whole fix, and it takes one tap.

What it learns The correction adjusts how that sender and that kind of message get weighed next time. The same miss doesn't repeat with the same person.

See also Trust & security, Triage, defined

The draft is right but the tone is off

Looks like The content of a drafted reply is accurate. It just doesn't sound like something you'd actually send to that specific person.

Why it happens Drafts are written from how you've corresponded with that person before. A new relationship doesn't have a sample to draw on yet. Neither does a new kind of conversation with an old one.

What contains it Nothing sends without your one-tap approval. A wrong tone you catch before sending costs you only the seconds it took to read the draft. It doesn't cost you an email someone else received.

What you do Edit it and send that instead.

What it learns Your edit becomes the sample it studies for that relationship. The next draft to that person starts closer to right.

See also How it works, In your voice, defined

It escalated something you would rather it had just handled

Looks like A routine confirmation lands in your queue asking for approval instead of going out quietly.

Why it happens Escalation exists to catch anything ambiguous or high-stakes. A cautious bar occasionally flags something you'd have been fine letting go.

What contains it Every escalation carries its own plain-language "why this reached you." The read is fast even when it wasn't strictly necessary. Nothing waits longer than your next glance at your drafts folder.

What you do Approve it, and turn the autonomy dial up a notch for that category if the pattern keeps showing up.

What it learns That category's threshold moves with the dial. The same kind of confirmation stops reaching you, once you've told it to stop.

See also Escalation, defined, Autonomy ladder, defined

It escalated too much and the interruptions stopped being useful

Looks like Enough approvals show up that they start blending together instead of standing out.

Why it happens This is what the early days look like. Beemy hasn't yet learned where your specific bar sits for what's worth interrupting you for.

What contains it Every escalation still carries its own reason, so the pattern is visible rather than a wall of undifferentiated pings.

What you do Tell it what class of thing to stop raising. Approve the pattern in bulk, or correct one and note why.

What it learns The correction moves the whole category, not just the one message it was attached to.

See also How it works, Escalation, defined

A draft leans on a detail from a thread that turned out to be stale

Looks like A reply references something that was true last week but has since changed.

Why it happens Drafts are written from the context Beemy has actually read. A change made somewhere it hasn't looked, like a phone call or a conversation in person, isn't visible to it.

What contains it The draft still waits in your drafts folder for approval. The one read you'd give it anyway is exactly where a stale detail gets caught.

What you do Fix the detail and send.

What it learns Nothing generalises from this one. The fact changed outside where it can see, so the fix is specific to that message, not a pattern to correct.

See also How it works

It called a calendar conflict that wasn't one, or missed one that was

Looks like Two events look like a clash and one was already declined, or a real overlap didn't get flagged at all.

Why it happens Conflict-checking reasons from what your Google Calendar actually shows. An informal hold made outside it, or a status change it hasn't caught up on yet, isn't information it has.

What contains it A conflict gets flagged for you, never resolved on its own: nothing is moved, accepted, or declined without your approval. A false flag costs one glance to dismiss.

What you do Dismiss the flag, or point out the one it missed.

What it learns The correction teaches it more about how you actually use your calendar. It learns which holds you treat as tentative, and which you treat as fixed.

See also Trust & security

You are worried a work reply could surface something personal

Looks like Not something that has happened. It's the worry itself, and a fair one to hold before handing over access to both parts of your inbox.

Why it happens It would be a real risk if Beemy kept one shared memory that any draft could pull from. It doesn't. Personal, work, and (on a team) shared context live in separate stores with no path between them.

What contains it This is architecture, not judgement. There's no route for a personal detail to reach a work draft. So there's nothing for a bad judgement call to leak in the first place.

What you do Nothing required. See /trust if you want the boundary stated in full, not just asserted here.

What it learns Nothing: this isn't a correction loop, it's a structural fact that doesn't change with use.

See also Trust & security, Silo, defined

A connected account went down or lost authorisation mid-day

Looks like Gmail or Google Calendar access lapses, whether from a revoked token or a provider outage. New mail then stops getting triaged.

Why it happens Beemy only sees what it has access to. If that access is revoked, or the provider it depends on is down, there's nothing new for it to read.

What contains it It fails closed, not open. It stops triaging silently, rather than guessing from stale data or acting on an assumption about what arrived while it was disconnected.

What you do Reconnect the account. Nothing sent before the gap gets duplicated or re-sent once it's back.

What it learns Nothing: this is an availability question, not a judgement one, and reconnecting is the entire fix.

See also Trust & security

One item, shown

What an escalation looks like.

Illustrative

Illustrative example, built with mock data.

In Gmail · work

What arrived

Client contact

Can you loop in their legal team on this?

A client asks to loop in a name that has never appeared on this thread. This is about a message on contract terms.

Beemy's message

Nothing is drafted. A plain "why this reached you" note explains it instead. An unrecognised name appears on a contract thread, so nothing gets guessed.

You read the reason and reply yourself. Nothing left the account on its own.

What we don't claim

No accuracy number, on purpose.

We register every third-party statistic we quote at The Numbers. That includes the ones we stopped quoting once we couldn't re-verify them. A self-reported accuracy figure from a private beta would fail that same bar, so we're not publishing one. What actually ships, and when, is tracked in status lanes on the roadmap, not hidden behind a number we can't back up.

Found one we didn't list

Tell us. That's how this list gets better.

If Beemy got something wrong in a way that isn't covered above, contact us and describe what happened. Real reports are how the catalogue grows and how the triage that missed it gets fixed.

Get your time back.

Beemy is in private beta. Join the waitlist and go from connect to a quiet, triaged inbox before you close your laptop tonight.

Private beta. No spam, ever.

Talk to us

Last updated